Storage-Aware Caching: Revisiting Caching for Heterogeneous Systems

Storage-Aware Caching:Revisiting Caching for Heterogeneous SystemsBrian ForneyAndrea Arpaci-Dusseau Remzi Arpaci-Dusseau

Wisconsin Network DisksUniversity of Wisconsin at Madison

Circa 1970Intense research period in policiesWide variety developed; many used todayExamples: Clock, LRUSimple storage environmentFocus: workloadAssumption: consistent retrieval costOSpolicybuffer cacheAppAppApp

TodayRich storage environmentDevices attached in many waysMore devicesIncreased device sophisticationMismatch: Need to reevaluateLANWANpolicybuffer cacheAppAppApp

Problem illustrationUniform workloadTwo disksLRU policy

Slow disk is bottleneckProblem: Policy is obliviousDoes not filter wellfastslow

General solutionIntegrate workload and device performanceBalance work across devicesWork: cumulative delayCannot throw out:Existing non-cost aware policy researchExisting caching software

Our solution: OverviewGeneric partitioning frameworkOld ideaOne-to-one mapping: device partitionEach partition has cost-oblivious policyAdjust partition sizesAdvantagesAggregates performance informationEasily and quickly adapts to workload and device performance changesIntegrates well with existing softwareKey: How to pick the partition sizes

OutlineMotivationSolution overviewTaxonomyDynamic partitioning algorithmEvaluationSummary

Partitioning algorithmsStaticPro: SimpleCon: WastefulDynamicAdapts to workloadHotspotsAccess pattern changesHandles device performance faultsUsed dynamic

Our algorithm: OverviewObserve: Determine per-device cumulative delayAct: Repartition cacheSave & reset: Clear last W requestsObserveActSave & reset

Algorithm: ObserveWant accurate system balance viewRecord per-device cumulative delay for last W completed disk requestsAt clientIncludes network time

Algorithm: ActCategorize each partitionPage consumersCumulative delay above threshold possible bottleneckPage suppliersCumulative delay below mean lose pages without decreasing performanceNeitherAlways have page suppliers if there are page consumersPage consumerPage supplierNeitherPage consumerPage supplierNeitherBeforeAfter

Page consumersHow many pages? Depends on state:WarmingCumulative delay increasingAggressively add pages: reduce queuingWarmCumulative delay constantConservatively add pagesCoolingCumulative delay decreasingDo nothing; naturally decreases

Dynamic partitioningEagerImmediately change partition sizesPro: Matches observationCon: Some pages temporarily unusedLazyChange partition sizes on demandPro: Easier to implementCon: May cause over correction

OutlineMotivationSolution overviewTaxonomyDynamic partitioning algorithmEvaluationSummary

Evaluation methodologySimulator

Workloads: synthetic and webCaching, RAID-0 clientLogGP network With endpoint contention Disks 16 IBM 9LZX First-order model: queuing, seek time & rotational time

Evaluation methodologyIntroduced performance heterogeneityDisk agingUsed current technology trendsSeek and rotation: 10% decrease/yearBandwidth: 40% increase/yearScenariosSingle disk degradation: Single disk, multiple agesIncremental upgrades: Multiple disks, two agesFault injectionUnderstand dynamic device performance change and device sharing effectsTalk only shows single disk degradation

Evaluated policiesCost-oblivious: LRU, ClockStorage-aware: Eager LRU, Lazy LRU, Lazy Clock (Clock-Lottery)Comparison: LANDLORDCost-aware, non-partitioned LRUSame as web caching algorithmIntegration problems with modern OSes

SyntheticWorkload: Read requests, exponentially distributed around 34 KB, uniform load across disks

A single slow disk greatly impacts performance.Eager LRU, Lazy LRU, Lazy Clock, and LANDLORD robust as slow disk performance degrades

Synthetic (LRU-based)

22.31830221.47918421.6329322.501009911.8902678

22.15429221.271055121.43257221.047283611.0763314

21.97150321.154575721.23907218.81581189.87784001

21.76354120.981381320.95034916.7363888.77347532

21.56213620.815501820.68799714.84021877.77015061

21.40541320.567296620.40514613.0676916.85042367

21.17965820.398797820.07844311.45553136.00395203

20.9630120.190019119.6754769.971266565.22351135

20.71843319.932220519.2284858.612929344.51369718

20.43517619.65090418.7800217.37426793.86311654

20.05864219.392981618.1854286.251097953.27377816

Eager LRU

LANDLORD

Lazy LRU

LRU

No cache

Age of slow disk (years)

Throughput (MB/s)

LRU-based

Synthetic (data)

Synthetic (LRU)

AgeEager LRULANDLORDLazy LRULRUNo cache

022.31830221.47918421.6329322.501009911.8902678

122.15429221.271055121.43257221.047283611.0763314

221.97150321.154575721.23907218.81581189.87784001

321.76354120.981381320.95034916.7363888.77347532

421.56213620.815501820.68799714.84021877.77015061

521.40541320.567296620.40514613.0676916.85042367

621.17965820.398797820.07844311.45553136.00395203

720.9630120.190019119.6754769.971266565.22351135

820.71843319.932220519.2284858.612929344.51369718

920.43517619.65090418.7800217.37426793.86311654

1020.05864219.392981618.1854286.251097953.27377816

Synthetic (Clock)

AgeClockLazy ClockNo cache

022.488704121.63102711.8902678

121.0522121.44133511.0763314

218.805337121.185649.87784001

316.728409520.9575718.77347532

414.825435120.7419057.77015061

513.064238420.4308016.85042367

611.444032920.1154796.00395203

79.9719685519.7752525.22351135

88.6214641419.2816574.51369718

97.3756685218.8016173.86311654

106.2531140918.2024163.27377816

Synthetic (Clock-based)

22.488704121.63102711.8902678

21.0522121.44133511.0763314

18.805337121.185649.87784001

16.728409520.9575718.77347532

14.825435120.7419057.77015061

13.064238420.4308016.85042367

11.444032920.1154796.00395203

9.9719685519.7752525.22351135

8.6214641419.2816574.51369718

7.3756685218.8016173.86311654

6.2531140918.2024163.27377816

Clock

Lazy Clock

No cache

Age of slow disk (years)

Throughput (MB/s)

Clock-based

Web (LRU-based)

160.646553146.6758098157.148103162.433180182.4647966904

159.413287145.9003505156.81857162.257253481.8391594405

155.34802144.1883992152.938273156.136846979.5502319365

144.362849137.6185628149.547187136.293345571.1343117637

143.615284130.4578955144.409007110.943717457.1995222751

146.285649134.8253256128.15222288.7434986744.7158626754

143.88189136.6751155102.72671969.3020919334.2355292989

139.413223134.375733290.50843851.7542978225.877293117

133.607377129.566885476.54555838.3567103419.3790536045

128.38637123.348024693.6412328.6910329114.3965840292

116.555625118.610126271.23063221.116853210.6126976973

Eager LRU

LANDLORD

Lazy LRU

LRU

No cache

Age of slow disk (years)

Throughput (MB/s)

Web (data)

Web (LRU-based)

AgeEager LRULANDLORDLazy LRULRUNo cache

0160.646553146.6758098157.148103162.433180182.4647966904

1159.413287145.9003505156.81857162.257253481.8391594405

2155.34802144.1883992152.938273156.136846979.5502319365

3144.362849137.6185628149.547187136.293345571.1343117637

4143.615284130.4578955144.409007110.943717457.1995222751

5146.285649134.8253256128.15222288.7434986744.7158626754

6143.88189136.6751155102.72671969.3020919334.2355292989

7139.413223134.375733290.50843851.7542978225.877293117

8133.607377129.566885476.54555838.3567103419.3790536045

9128.38637123.348024693.6412328.6910329114.3965840292

10116.555625118.610126271.23063221.116853210.6126976973

Synthetic (LRU-based)

22.31830221.47918421.6329322.501009911.8902678

22.15429221.271055121.43257221.047283611.0763314

21.97150321.154575721.23907218.81581189.87784001

21.76354120.981381320.95034916.7363888.77347532

21.56213620.815501820.68799714.84021877.77015061

21.40541320.567296620.40514613.0676916.85042367

21.17965820.398797820.07844311.45553136.00395203

20.9630120.190019119.6754769.971266565.22351135

20.71843319.932220519.2284858.612929344.51369718

20.43517619.65090418.7800217.37426793.86311654

20.05864219.392981618.1854286.251097953.27377816

Eager LRU

LANDLORD

Lazy LRU

LRU

No cache

Age of slow disk (years)

Throughput (MB/s)

LRU-based

Synthetic (data)

Synthetic (LRU)

AgeEager LRULANDLORDLazy LRULRUNo cache

022.31830221.47918421.6329322.501009911.8902678

122.15429221.271055121.43257221.047283611.0763314

221.97150321.154575721.23907218.81581189.87784001

321.76354120.981381320.95034916.7363888.77347532

421.56213620.815501820.68799714.84021877.77015061

521.40541320.567296620.40514613.0676916.85042367

621.17965820.398797820.07844311.45553136.00395203

720.9630120.190019119.6754769.971266565.22351135

820.71843319.932220519.2284858.612929344.51369718

920.43517619.65090418.7800217.37426793.86311654

1020.05864219.392981618.1854286.251097953.27377816

Synthetic (Clock)

AgeClockLazy ClockNo cache

022.488704121.63102711.8902678

121.0522121.44133511.0763314

218.805337121.185649.87784001

316.728409520.9575718.77347532

414.825435120.7419057.77015061

513.064238420.4308016.85042367

611.444032920.1154796.00395203

79.9719685519.7752525.22351135

88.6214641419.2816574.51369718

97.3756685218.8016173.86311654

106.2531140918.2024163.27377816

Synthetic (Cloc